My experience is that around 200-300 lines, the LLM is no longer able to understand the context of its own work. At that point, it also usually contains a significant mistake.

And around that scale, you’ll get much better code if you refactor it yourself and clean up its logic (which tends to be verbose, at best).

Someday, it will be better; today, none of them (Grok, ChatGPT, Claude) are good enough for what you describe.